The recording proves what they really said: leaked audio of a public figure is authentic, because you can hear their own voice saying it
By The Conspiratory EditorsAugust 20, 2026
Where the evidence lands: ContradictedThat a circulating audio recording, or the audio track of a circulating video, is authentic evidence of what a public figure said, on the grounds that the voice is recognisably theirs and the accompanying footage can be verified as real.
Contradicted. Treating a recognisable voice as proof of authorship stopped working somewhere around 2023, and the habit has not caught up. What is rated here is that inference, not any particular clip. The documented record is unambiguous: an AI-generated Biden voice really did telephone New Hampshire voters before the 2024 primary, a manipulated recording of a Slovak party leader really did circulate two days before that country's election, and in August 2026 two separate fact-checkers found fabricated audio laid over authentic video of a refugee in Tunisia and of the Indian prime minister. In each case the picture, the person and often the event were genuine, and only the sound was invented. That combination defeats the check most people have been taught to run, because asking whether the video is real returns yes. Note also what the file does not claim: that any of these recordings decided anything. A Harvard Kennedy School analysis of the Slovak case argues the deepfake's electoral effect has been overstated, and that is included here rather than left out.
Believed by: Everyone, at least briefly, which is the difficulty. Fabricated audio has been shared by ordinary partisans, by mainstream accounts, and in at least one case by people who thought they were warning others about the technology
The first artifact logged here since this file was built, and the technique is arriving faster than the file's own framing assumed. Fact Crescendo rated False on 24 August a twenty-second clip circulating in India as the actor Akshay Kumar mocking the Declassified: Operation Sindoor documentary and the Indian Army's performance. The picture is authentic and old, traced to a December 2025 recording; the audio was digitally manipulated. The documentary itself is real, a two-part Warner Bros. Discovery production that premiered on Discovery Channel India on 15 August 2026 featuring the National Security Adviser, the Chief of Defence Staff and the Army Chief. Two things are worth recording. The clip did not have to establish its own context, because a genuine news event nine days earlier had already supplied it; a viewer encountering the video already knew what documentary was meant. And the post came with a caption spelling out what the sarcasm was supposed to mean, which is a tell in itself: fabricated audio that needed no explanation would not come with one. Note also a scoping drift worth correcting. Three audio-substitution cases have been logged on this site in the past week, an authentic clip of a refugee given fabricated English speech, an edit adding a national anthem behind a head of government, and now this, and two of them were filed against the doctored-images file rather than here. They belong together, because they share the property that makes them hard: the image is genuine, the person really is present, and the only fabricated element is the one a viewer cannot see. source →
A second fact-checker reached the same finding on the Modi clip one day later, which matters for a claim of this kind. Factly published on 20 August, in Telugu, that the video circulating with the claim that the prime minister walked while the national anthem was being played has been edited. Fact Crescendo had reported on 19 August that the anthem is not playing in the original footage. Two independent Indian outlets, one day apart, describing the same manipulation from opposite ends: one naming what was added, the other naming what the addition was supposed to prove. Corroboration is worth more here than it would be for a fabricated image, because a soundtrack leaves so little for a reader to inspect that a single outlet's finding is harder to sanity-check independently. source →
The full story
The voice used to settle it
For about a century, a recording was the end of an argument. Documents could be forged and witnesses could be mistaken, but if you could hear somebody saying the thing in their own voice, the question of whether they had said it was closed. Watergate turned on tapes. So did a long line of scandals after it.
That property is gone, and it went quietly. There was no single moment when voice stopped functioning as identification, which is why the habit of trusting it has outlived the technical fact underneath it.
This file is not about any one clip. It is about the inference: that hearing somebody say something is evidence they said it. Four documented cases across three years are enough to show what that inference now costs.
Why people keep trusting recordings
The instinct deserves more respect than it usually gets. It is not gullibility, it is an old rule that used to be correct, and people are applying it in good faith. Anyone over about thirty formed their sense of what counts as proof in a period when faking a voice convincingly was genuinely hard.
There is also less to inspect. A suspicious photograph offers hands, shadows, reflections, edges: a dozen places for the eye to catch. A voice offers almost nothing. On a phone line, compressed and brief, even a poor fake has very little surface for scepticism to grip.
And it is worth conceding the strongest version of the sceptical case, which is that the panic can be overdone. The Slovak recording is regularly described as the first deepfake to swing an election. A commentary in the Harvard Kennedy School Misinformation Review argues that account is too simple, and that a receptive electorate, encrypted messaging apps and politicians amplifying the material did much of the work. If the technology is being credited with power it does not have, saying so is part of getting this right.
Four cases, three years, one technique
Slovakia, 28 September 2023.Two days before the parliamentary election, audio spread that appeared to catch Progressive Slovakia's chairman and a Dennik N journalist discussing rigging the vote. Both said at once it was fabricated. AFP debunked it the same day, citing poor voice quality, unintelligible passages and unnatural speech patterns.
New Hampshire, January 2024. Thousands of households answered the phone to an AI-generated Joe Biden telling them to skip the primary and save their vote for November. The calls were spoofed to look like they came from a local Democrat.
Tunisia footage, August 2026. Full Fact traced a clip of a migrant apparently demanding money and a house in English to Associated Press footage from February 2024 of a Sudanese refugee speaking Arabic about safety in Albania. We cover this one in the file on recycled footage.
India, August 2026. Fact Crescendo found that a video circulating as proof the prime minister had disrespected the national anthem was altered. The anthem is not playing in the original. It was added.
The last two are the ones that change the problem. In both, the footage is authentic, the person really was present, and the event really happened. Only the sound is invented.
The check everyone was taught to run
Three years of public education about synthetic media have converged on a single question: is this video real? Look at the hands. Reverse image search it. Ask whether the scene was generated.
Run that on the Modi clip and it comes back clean, because the clip is real. Run it on the Tunisia footage and you land on Associated Press material from 2024, which is about as good a provenance result as anyone could hope for. The verification succeeds and returns the wrong answer.
This is what makes the audio variant different in kind from a generated video rather than just cheaper. A wholly synthetic clip fails a provenance check. A genuine clip with a fabricated soundtrack passes one, and hands the person who ran it a feeling of having confirmed something.
The detection resources are also lopsided. Image provenance tooling is widespread and improving. Audio forensics is specialised, and every 2026 case here was caught by a fact-checker doing patient sourcing work, tracing a clip back to its original, rather than by any instrument.
The one time the law was tested end to end
The New Hampshire robocalls are the only case in this file that ran the whole length of a legal system, and the result is not what either side would have predicted.
The FCC ruled in February 2024 that AI-generated voices in robocalls are illegal under existing law. That September it levied a six million dollar fine against the political operative behind the calls, and the telecom carrier that transmitted them settled separately for a million dollars and a compliance programme. New Hampshire went on to require greater transparency in AI-generated political advertising.
Then, on 13 June 2025, a Belknap County jury acquitted him on every count. He had faced 11 felony counts of voter suppression and 11 of impersonating a political candidate, consolidated from charges brought by several county attorneys. He had admitted making the calls. He testified that he did it as a warning about the dangers of artificial intelligence. The state attorney general thanked the jury and said his office would keep enforcing election law.
So the same admitted conduct produced a substantial civil penalty and a complete criminal acquittal. Whatever else that shows, it means the honest answer to “is this illegal?” is more complicated than either a headline fine or a headline acquittal suggests, and anyone citing one half of it is telling half the story.
A harder question than is it real
The practical upshot is uncomfortable, because the replacement for the old rule is slower than the rule was.
Asking whether a recording is authentic no longer has a quick answer. What has an answer is where it came from: who first posted it, whether the footage exists anywhere else, and whether the original carries the same soundtrack. Every case here was resolved that way, by somebody finding the source rather than by analysing the artifact. Full Fact matched a clip to Associated Press footage from 2024. Fact Crescendo found the original without the anthem.
It is also worth being clear about what remains unproven. Fabricated audio demonstrably circulates, demonstrably reaches large audiences, and has demonstrably targeted heads of government and national elections. That it changes outcomes is a separate claim, and the most careful published analysis of the best-known case argues it has been overstated. Both of those things can be true, and this file holds them together rather than picking the more dramatic one.
What's still unexplained
- What actually deters it? The one full test in the record produced a 6 million dollar civil fine and an acquittal on all 22 criminal counts, arising from the same admitted conduct. Whether that outcome reflects a gap in the statutes, the specific facts of an unsanctioned primary, or a jury's view of intent is not settled, and it leaves the deterrence question genuinely open.
- How much does fabricated audio actually change? The Slovak case is routinely cited as the first election influenced by a deepfake, and a peer-reviewed commentary argues that reading is too simple. Nobody has a clean method for separating the effect of a recording from the effect of an electorate already prepared to believe it.
- Who checks audio outside well-funded newsrooms? Image provenance tools are widely available and improving. Audio forensics is more specialised, and most of the 2026 instances were caught by regional fact-checkers doing careful sourcing work rather than by any tool.
- Does labelling reach this at all? The EU AI Act's transparency obligations applied from 2 August 2026 and require disclosure that a person can understand without a detection tool. A fabricated soundtrack laid over genuine footage by an anonymous account is exactly the case where there is no provider or deployer to place the obligation on.
Point by point
The claim: It is unmistakably their voice, so it must be them.
What the record shows: Voice is now among the cheapest things to synthesise. The New Hampshire robocall was, on the account given to NBC News, produced by a street magician who was hired to do it. The technical barrier that made a recognisable voice function as identification for a century is gone, and no amount of familiarity with how somebody sounds substitutes for provenance.
The claim: The video is real, so the audio must be too.
What the record shows: This is the assumption the 2026 cases break, and it is the reason they matter more than a straightforward deepfake. In both the Tunisia and the Modi clips the footage is genuine and the person really was there. Only the soundtrack is invented. Every reverse image search, every provenance check on the picture, and every instinct about whether the scene looks generated returns a clean result, because the picture is not the lie.
The claim: Fabricated audio is obvious once you listen carefully.
What the record shows: Sometimes, and that is what saved the Slovak case in part: AFP identified poor voice quality, unintelligible passages and unnatural speech patterns. But those were the tells of a 2023 system, and the resources to check them are unevenly distributed. A regional fact-checker with an audio forensics budget is rarer than a viral clip.
The claim: If the law forbids it, the problem is largely handled.
What the record shows: The record cuts both ways here and the file states it plainly. The FCC ruled AI voices in robocalls illegal in February 2024 and levied a 6 million dollar fine that September, and the carrier settled for a further million. Yet a jury acquitted the man who admitted making them on all 22 counts. Civil and criminal enforcement reached opposite conclusions about the same admitted act.
The claim: A deepfake recording can swing a national election.
What the record shows: This is the claim the file is most careful about, because overstating it is its own distortion. The Harvard Kennedy School Misinformation Review commentary on Slovakia argues precisely that the standard account is too simple, and that pre-existing receptivity to pro-Russian messaging, encrypted messaging apps and amplification by politicians did much of the work usually attributed to the audio. Fabricated recordings are real and demonstrably circulate; that they decide outcomes is not established.
The claim: The timing of these releases is incidental.
What the record shows: Not in the clearest case. The Slovak recording appeared during a 48-hour pre-election moratorium in which media outlets and politicians are expected to stay silent, which is precisely the window in which a correction is hardest to distribute. Whether or not that timing was chosen, its effect was to release the claim while the machinery for answering it was switched off.
Timeline
- 2023-09-28Two days before Slovakia's parliamentary election, an audio recording spreads on Facebook, TikTok and YouTube appearing to capture Progressive Slovakia chairman Michal Simecka and Dennik N journalist Monika Todova discussing rigging the vote. Both immediately say it is fabricated. AFP debunks it the same day, describing a recording with numerous signs of manipulation, including poor voice quality, unintelligible passages and unnatural speech patterns. It is released during the 48-hour pre-election moratorium, when Slovak media outlets and politicians are expected to stay silent, which limits how widely it can be answered.
- 2024-01Thousands of New Hampshire households receive robocalls carrying an AI-generated voice resembling President Joe Biden, urging them to skip the presidential primary and save their vote for November. The calls are spoofed to appear to come from a prominent local Democrat.
- 2024-02The Federal Communications Commission rules that the use of AI-generated voices in robocalls is illegal under existing law.
- 2024-02A street magician tells NBC News he was hired by the political operative Steven Kramer to produce the artificial Biden recording. Kramer had previously worked for the campaign of Dean Phillips, who was challenging Biden for the nomination; the Phillips campaign denies involvement and distances itself.
- 2024-08-22The Harvard Kennedy School Misinformation Review publishes a commentary complicating the account of Slovakia as the first election swung by a deepfake, pointing instead to pre-existing susceptibility to pro-Russian messaging, the role of encrypted messaging apps, misinformation effects in low-trust environments, and amplification by politicians themselves.
- 2024-09The FCC levies a 6 million dollar fine against Kramer. Lingo Telecom, the voice service provider that carried the calls, separately agrees to pay 1 million dollars and adopt a compliance programme. New Hampshire subsequently enacts a law requiring greater transparency in AI-generated political advertising.
- 2025-06-13A Belknap County Superior Court jury acquits Kramer on all charges. Cases brought by several county attorneys had been consolidated into a single trial in which he faced 11 felony counts of voter suppression and 11 counts of impersonating a political candidate. He had admitted creating the robocalls and testified in Laconia that he orchestrated them as a warning about the dangers of artificial intelligence. Attorney General John Formella thanked the jury and said his office remained committed to enforcing election law.
- 2026-08-18Full Fact reports that a clip shared thousands of times, appearing to show a migrant saying in English that he wants money, a free house and an English woman, is altered. The source is Associated Press footage from February 2024 of a Sudanese refugee in Tunisia speaking Arabic about safety in Albania. Beyond the fabricated audio, the man's mouth and gestures were changed and so were the positions of the other men around him.
- 2026-08-19Fact Crescendo reports that a video circulating as evidence that Indian Prime Minister Narendra Modi disrespected the national anthem is altered: the anthem is not playing in the original footage. Rated Misleading. Two continents, two fact-checkers, two days.
Contradicted. Treating a recognisable voice as proof of authorship stopped working somewhere around 2023, and the habit has not caught up. What is rated here is that inference, not any particular clip. The documented record is unambiguous: an AI-generated Biden voice really did telephone New Hampshire voters before the 2024 primary, a manipulated recording of a Slovak party leader really did circulate two days before that country's election, and in August 2026 two separate fact-checkers found fabricated audio laid over authentic video of a refugee in Tunisia and of the Indian prime minister. In each case the picture, the person and often the event were genuine, and only the sound was invented. That combination defeats the check most people have been taught to run, because asking whether the video is real returns yes. Note also what the file does not claim: that any of these recordings decided anything. A Harvard Kennedy School analysis of the Slovak case argues the deepfake's electoral effect has been overstated, and that is included here rather than left out.
Reviewed by The Conspiratory Editors · Last reviewed August 25, 2026 · How we rate
Common questions
Is Fabricated audio true?
Contradicted. Treating a recognisable voice as proof of authorship stopped working somewhere around 2023, and the habit has not caught up. What is rated here is that inference, not any particular clip. The documented record is unambiguous: an AI-generated Biden voice really did telephone New Hampshire voters before the 2024 primary, a manipulated recording of a Slovak party leader really did circulate two days before that country's election, and in August 2026 two separate fact-checkers found fabricated audio laid over authentic video of a refugee in Tunisia and of the Indian prime minister. In each case the picture, the person and often the event were genuine, and only the sound was invented. That combination defeats the check most people have been taught to run, because asking whether the video is real returns yes. Note also what the file does not claim: that any of these recordings decided anything. A Harvard Kennedy School analysis of the Slovak case argues the deepfake's electoral effect has been overstated, and that is included here rather than left out.
What is Fabricated audio?
For most of the last century a recording of somebody's voice was close to conclusive. Tape could be spliced and quotes could be trimmed, but the voice itself was the part nobody could fake. That is no longer true, and the change arrived faster than the habit of trusting it could adjust. Two days before Slovakia's 2023 parliamentary elect…
What does the evidence show?
Voice is now among the cheapest things to synthesise. The New Hampshire robocall was, on the account given to NBC News, produced by a street magician who was hired to do it. The technical barrier that made a recognisable voice function as identification for a century is gone, and no amount of familiarity with how somebody sounds substitu…
Why do people believe it?
The evidentiary instinct is old, reasonable and was correct until very recently. Recordings really were hard to fake convincingly, and a century of journalism, litigation and scandal was built on that being true. People are not being credulous; they are applying a rule that used to work.
What is still unresolved?
What actually deters it? The one full test in the record produced a 6 million dollar civil fine and an acquittal on all 22 criminal counts, arising from the same admitted conduct. Whether that outcome reflects a gap in the statutes, the specific facts of an unsanctioned primary, or a jury's view of intent is not settled, and it leaves th…
Sources
- 1.Political operative who admitted to creating fake Biden robocalls found not guilty, New Hampshire Public Radio (2025)
- 2.New Hampshire jury acquits consultant behind AI robocalls mimicking Biden on all charges, Courthouse News Service (2025)
- 3.Criminal charges and FCC fines issued for deepfake Biden robocalls, NPR (2024)
- 4.Beyond the deepfake hype: AI, democracy, and the Slovak case, Harvard Kennedy School Misinformation Review (2024)
- 5.Video of migrant saying he wants money, free house and English woman has been edited, Full Fact (2026)
- 6.Edited video goes viral as PM Modi disrespects the national anthem, Fact Crescendo (2026)
Embed this verdict on your site
Paste this snippet to show our sourced verdict as a small card, with a link back to the full case file. Free to use.
<iframe src="https://theconspiratory.com/embed/fake-audio-recordings" title="The Conspiratory verdict" width="520" height="190" style="border:0;max-width:100%" loading="lazy"></iframe>Help us investigate
This is a living case file. If you spot an error or know evidence we missed, tell us, and weigh in on where you land.
Where do you land?
Cast your read on this one.
Comments
Add your take. Comments are read and approved by a human before they appear, so keep it on topic and civil. Please do not accuse named, living people of crimes.